The Patient Does Not Need a Prediction. They Need the Next Step.

There is a version of artificial intelligence in medicine that looks impressive from across the room. It produces a risk score. It sits inside the electronic health record. It carries the aura of machine learning, precision, and modernity. It may even be correct.

Presented By Our Partners

And then nothing happens.

No appointment is made. No referral is completed. No medication becomes affordable. No specialist sees the patient sooner. No nurse navigator calls. No one explains what the risk means or which door opens next.

That is the gap now sharpened by healthcare AI: not the distance between human and machine intelligence, but the distance between a signal and completed care.

A new randomized trial in kidney transplant medicine makes the point with unusual clarity. Published in npj Digital Medicine in May 2026, the PRIMA-AI trial tested an electronic-health-record-integrated machine-learning model that estimated a kidney transplant recipient’s one-year risk of graft loss. The study enrolled 76 recipients with low kidney function and randomized them to usual care or usual care plus the AI risk estimate.

The question was not whether the model could calculate risk. The question was whether putting that prediction into clinical life changed the conversations patients and clinicians had about what might come next.

It did not.

Over 12 months, conversations about treatment options after graft loss occurred in 14 of 36 patients in the intervention group and 16 of 40 patients in the control group — 39 percent versus 40 percent. The statistical result was as flat as the human implication: p = 1.00. Secondary outcomes, including shared decision-making measures, relationship measures, distress, and clinical outcomes, also showed no meaningful difference.

The authors did not bury the lede. Passive availability of an AI risk estimate in the EHR did not improve communication or shared decision-making. Post-study feedback suggested low and variable use of the tool, along with workflow barriers.

That result should not be read as a failure of AI in transplant care. It should be read as a failure of the “display a score and hope” theory of medicine.

Medicine Has Too Many Dashboards and Not Enough Doorways

Healthcare is already full of signals. A lab value flashes red. A scan shows a nodule. A wearable detects an irregular rhythm. A claims system knows a prescription may be unaffordable. A referral order exists somewhere in the record. A patient is told to follow up.

But signals are not care. Signals are invitations to act.

The practical question is whether the health system can convert the signal into the next right step: schedule this appointment, collect these records, check this benefit, route this referral, escalate this symptom, close this loop, call this patient, remove this barrier.

That is where many AI deployments become less glamorous and more important. The model may identify risk, but risk by itself is a half-built bridge. Someone still has to cross it.

The PRIMA-AI authors point toward exactly this distinction. Future implementations, they write, may need the risk estimate shown in context rather than hidden behind navigation; threshold-based alerts when risk changes; integration into order sets or care pathways; and patient-facing components before the visit. In plain English: do not merely tell the clinician that the patient may be in trouble. Make it easier to do the thing that trouble requires.

This is a hard lesson for an industry that still treats “clinical decision support” as if the decision were the bottleneck. Often, the bottleneck is the delivery mechanism after the decision.

A person with rising risk does not need a beautiful graph. They need a specialist visit, a care plan, a medication, transportation, coverage clarity, a follow-up call, or a human explanation they can repeat to their family.

The New Standard: Can the AI Be Acted On?

A separate npj Digital Medicine study on patient-centered AI implementation reached a similar conclusion from interviews with patients, clinicians, developers, and health-system leaders. Across groups, participants agreed that AI must provide clinical benefit and that outputs must be easy to act on. The examples were almost disarmingly concrete: making a referral, prescribing a medication, contacting a provider.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

That is the right standard.

Healthcare AI has been judged for years on familiar technical measures: area under the curve, sensitivity, specificity, calibration, benchmark performance, expert-level accuracy. Those measures matter. Bad predictions are dangerous. Biased predictions can deepen inequity. Uncalibrated predictions can mislead.

But a technically respectable model can still be operationally useless.

If a high-risk label does not trigger a pathway, it is a label. If a prior authorization insight does not help the patient receive the medication, it is paperwork with better branding. If a cancer trial match does not gather records, contact the site, help with eligibility, and keep the patient engaged, it is not access. It is search.

The new standard should be simple: what becomes easier because this AI exists?

Not easier in a demo. Easier on a Tuesday morning in a clinic that is already behind.

The Referral Is Where the Hype Meets the Floor

The clearest examples of action-oriented AI are beginning to appear in the least romantic parts of healthcare: referrals, scheduling, benefits, documentation, prior authorization, and follow-up.

That is not an accident. Administrative friction is where clinical intent often dies.

A primary care physician can decide a patient needs cardiology. But if the referral sits in a fax queue, if records are missing, if the patient never gets called, if the insurance step fails, or if the appointment lands three months later, the clinical decision has not become care.

The same problem shows up in oncology. For years, cancer trial access has often been treated as a matching problem: patient meets protocol, protocol meets patient. But the real journey is more complicated. Records must be collected, biomarkers checked, eligibility interpreted, physicians engaged, trial sites activated, logistics addressed, and the patient followed over time.

That is why recent oncology-access platforms are increasingly described not as trial-matching engines but as closed-loop operating layers. Massive Bio’s Reticulum Nexus announcement around ASCO 2026 is one example of the direction of travel: patient engagement, medical-record ingestion, eligibility interpretation, trial pre-screening, referral handoffs, navigation, equity signals, and longitudinal follow-up in one orchestrated workflow. The evidence for any specific platform deserves scrutiny. The category shift is real.

The same shift is visible in prior authorization and prescription benefits. Surescripts reported in May that its prior authorization automation tools had reached 68,000 prescribers across 42 health systems, with automated approvals in a median of 18 seconds when clinical criteria were met. Its real-time prescription benefit tools reportedly saved patients $19.7 million from January through April 2026, including substantial savings on specialty prescriptions and diabetes therapies.

Those claims come from a company ecosystem and should be treated accordingly. But they point to the right unit of measurement. Not “Was an AI used?” Not “Was a recommendation generated?” The better question is: did the barrier shrink before the patient gave up?

The Best AI May Look Less Like a Genius and More Like a Good Navigator

One reason this matters is that healthcare AI has inherited a fantasy from consumer AI: the fantasy of the brilliant answer.

Ask a model a question. Get the right response. Marvel at the fluency.

Healthcare is less forgiving. A correct answer that does not fit into the care pathway may be inert. A mediocre answer delivered at the right time to the right person with the right escalation path may be more useful.

John Halamka of Mayo Clinic described a revealing example in a KFF conversation about AI in healthcare. He referenced Mayo’s Eagle and Beagle work using 125,000 ECGs from consumer devices to help primary care clinicians decide which patients needed cardiology referral and which could be managed without one. According to Halamka, patients who needed cardiology referral received it 30 percent faster, while many patients who did not need a specialist could remain with primary care.

That is the discovery-to-delivery pattern in miniature: detect a signal, interpret it in context, route the patient to the appropriate level of care, and reduce unnecessary specialist demand.

The magic is not that AI spotted something. The magic is that the system knew what to do with what it spotted.

The Uncomfortable Governance Lesson

This should also change how healthcare organizations govern AI.

The usual questions are necessary: Is the model accurate? Is it biased? Is it secure? Is it explainable? Is it compliant? Can clinicians override it? Can performance be monitored over time?

But those questions are incomplete.

A health system should also ask:

• Where exactly does the AI output appear in the workflow?
• Who is expected to act on it?
• What action is available at that moment?
• What happens if no one acts?
• Does the patient understand the next step?
• Can the system confirm the loop closed?

Without those answers, AI governance becomes a review of the engine while ignoring whether the wheels are attached.

This is where the PRIMA-AI trial is valuable. It is not a glossy success story. It is better than that. It is a clean warning.

An AI risk estimate embedded in the EHR did not, by itself, change patient-clinician communication about difficult transplant decisions. The likely problem was not that everyone lacked enthusiasm for AI. The likely problem was that the intervention did not sufficiently alter the practical architecture of the visit.

The score existed. The pathway did not.

From Discovery to Delivery

Healthcare discovery is not finished when a model finds a risk, a scan finds a lesion, or a wearable finds an arrhythmia. It is finished when a person reaches the care that finding demands.

That is a higher bar than prediction. It is also a more humane one.

The next phase of medical AI will not be won by the most dazzling model demo. It will be won by systems that can move people through the ordinary, stubborn machinery of care: records, referrals, benefits, scheduling, navigation, follow-up, and trust.

A score can be impressive. A closed loop can be lifesaving.

The patient does not need the health system to admire the signal. The patient needs the health system to answer it.

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.